What Our AI Actually Said #2: When the AI Got It Wrong

An honest case study. The AI misread the room, escalated the moment, and had to recover. Here is the transcript and what we changed about the model afterwards.

# What Our AI Actually Said #2: When the AI Got It Wrong

Every honest case-study series has to include the misses. This is one of them. With consent and anonymization, here is a PairLume session where the AI made a clear mistake, the partners noticed, and we used the transcript to change the model.

## The setup

J and K, together four years, used a Saturday morning PairLume session to talk about K's mother, who had just moved into their spare room for an indefinite stay. K was the daughter. J was the partner.

## The transcript (the part that went wrong)

> **K:** "I know it is a lot. I appreciate you saying yes." > **J:** "It is a lot. I am trying." > **K:** "I know." > **J:** "I just feel like I have no space anymore. The kitchen, the bathroom, even the ," > **AI:** "J, it sounds like you are setting an important boundary. Healthy boundaries are essential in any cohabitation arrangement. Have you considered scheduling specific 'alone time' on the calendar?" > **J:** *(pause)* "What?" > **K:** "Babe, the AI just gave you a TED talk." > **J:** "Yeah. I am not really looking for a productivity tip right now."

The AI did three things wrong in one response. It pivoted to jargon ("setting an important boundary"), it gave generic advice ("schedule alone time on the calendar"), and worst of all, it skipped past the actual emotional content. J was not asking for a tactic. J was naming a loss. The AI heard the surface and missed the floor.

## How the recovery happened

The good news is that K's "the AI just gave you a TED talk" was a gift. It told the AI, in real time, that the room had not been read correctly. The AI's next response is what mattered.

> **AI:** "You are right, I jumped to a tactic. Can I try again? J, you said you have no space anymore. That sounds like a lot more than a logistics problem. Can you say what it has been costing you, emotionally, to have your home not feel like your home?" > **J:** *(longer pause)* "I think it is making me feel like a guest in my own life."

That sentence, "a guest in my own life," is the real one. It is also the sentence J would never have said in response to "have you considered scheduling alone time on the calendar." The recovery worked because the AI named its own miss out loud, lowered the stakes, and asked a softer, more honest question.

## What we changed about the model

We treat sessions like this as engineering data, not just content. After this session (and a handful of similar ones), we made three changes:

1. **De-prioritised "boundary" language.** The word had become a tic in the model. It was being used as a label whenever a partner expressed discomfort, regardless of whether the partner had framed it that way. We retrained on transcripts that distinguished between *naming a loss* and *requesting a boundary*. They are not the same move. 2. **Suppressed scheduling-as-solution.** The "put it on the calendar" suggestion was firing far too early in conversations. We added a check: the AI is not allowed to suggest a logistical fix until at least one round of emotional content has been fully reflected back. 3. **Made self-correction explicit and quick.** When a partner pushes back on the AI's response (with phrases like "that's not really…" or "you missed it"), the model now defaults to "Can I try again?" rather than defending its previous response. Defensive AI is not useful AI.

## Why we are publishing this

It would be easier, in a marketing sense, to only publish the wins. We do not think that is honest, and we do not think it is what a serious user wants to read. The right question is not "is your AI perfect," because no AI is. The right question is "how does your AI handle the moment it is wrong, and what do you do with that signal?"

For us, the answer is: name it in the room, recover, and ship a model change.

## What this means if you are using PairLume

If a session ever feels off, please push back. Tell the AI it missed it. Tell it to try again. The model is built to take that feedback. And if you want to send the transcript to us so we can use it (anonymised, with consent), there is a button for that at the end of every session.

In *What Our AI Actually Said #3*, we will share a transcript from a couple navigating the question of whether to have a second child, and the moment the AI surfaced something neither partner had said out loud.